fix: closes #1277

docs/ai-tools-org-assessment.md

@@ -1,66 +0,0 @@
-# AI Tools Org Assessment — Implementation Tracker
-
-**Issue:** #1119
-**Research by:** Bezalel
-**Date:** 2026-04-07
-**Scope:** github.com/ai-tools — 205 repositories scanned
-
-## Summary
-
-The `ai-tools` GitHub org is a broad mirror/fork collection of 205 AI repos.
-~170 are media-generation tools with limited operational value for the fleet.
-7 tools are strongly relevant to our infrastructure, multi-agent orchestration,
-and sovereign compute goals.
-
-## Top 7 Recommendations
-
-### Priority 1 — Immediate
-
-- [ ] **edge-tts** — Free TTS fallback for Hermes (pip install edge-tts)
-  - Zero API key, uses Microsoft Edge online service
-  - Pair with local TTS (fish-speech/F5-TTS) for full sovereignty later
-  - Hermes integration: add as provider fallback in text_to_speech tool
-
-- [ ] **llama.cpp** — Standardize local inference across VPS nodes
-  - Already partially running on Alpha (127.0.0.1:11435)
-  - Serve Qwen2.5-7B-GGUF or similar for fast always-available inference
-  - Eliminate per-token cloud charges for batch workloads
-
-### Priority 2 — Short-term (2 weeks)
-
-- [ ] **A2A (Agent2Agent Protocol)** — Machine-native inter-agent comms
-  - Draft Agent Cards for each wizard (Bezalel, Ezra, Allegro, Timmy)
-  - Pilot: Ezra detects Gitea failure -> A2A delegates to Bezalel -> fix -> report back
-  - Framework-agnostic, Google-backed
-
-- [ ] **Llama Stack** — Unified LLM API abstraction layer
-  - Evaluate replacing direct provider integrations with Stack API
-  - Pilot with one low-risk tool (e.g., text summarization)
-
-### Priority 3 — Medium-term (1 month)
-
-- [ ] **bolt.new-any-llm** — Rapid internal tool prototyping
-  - Use for fleet health dashboard, Gitea PR queue visualizer
-  - Can point at local Ollama/llama.cpp for sovereign prototypes
-
-- [ ] **Swarm (OpenAI)** — Multi-agent pattern reference
-  - Don't deploy; extract design patterns (handoffs, routines, routing)
-  - Apply patterns to Hermes multi-agent architecture
-
-- [ ] **diagram-ai / diagrams** — Architecture documentation
-  - Supports Alexander's Master KT initiative
-  - `diagrams` (Python) for CLI/scripted, `diagram-ai` (React) for interactive
-
-## Skip List
-
-These categories are low-value for the fleet:
-- Image/video diffusion tools (~65 repos)
-- Colorization/restoration (~15 repos)
-- 3D reconstruction (~22 repos)
-- Face swap / deepfake tools
-- Music generation experiments
-
-## References
-
-- Issue: https://forge.alexanderwhitestone.com/Timmy_Foundation/the-nexus/issues/1119
-- Upstream org: https://github.com/ai-tools

mimo-swarm/scripts/mimo-dispatcher.py

@@ -7,6 +7,7 @@ routes to lanes, and spawns one-shot mimo-v2-pro workers.
 No new issues created. No duplicate claims. No bloat.
 """
 
+import glob
 import json
 import os
 import sys
@@ -38,6 +39,7 @@ else:
 
 CLAIM_TIMEOUT_MINUTES = 30
 CLAIM_LABEL = "mimo-claimed"
+MAX_QUEUE_DEPTH = 10  # Don't dispatch if queue already has this many prompts
 CLAIM_COMMENT = "/claim"
 DONE_COMMENT = "/done"
 ABANDON_COMMENT = "/abandon"
@@ -451,6 +453,13 @@ def dispatch(token):
     prefetch_pr_refs(target_repo, token)
     log(f"  Prefetched {len(_PR_REFS)} PR references")
 
+    # Check queue depth — don't pile up if workers haven't caught up
+    pending_prompts = len(glob.glob(os.path.join(STATE_DIR, "prompt-*.txt")))
+    if pending_prompts >= MAX_QUEUE_DEPTH:
+        log(f"  QUEUE THROTTLE: {pending_prompts} prompts pending (max {MAX_QUEUE_DEPTH}) — skipping dispatch")
+        save_state(state)
+        return 0
+
     # FOCUS MODE: scan only the focus repo. FIREHOSE: scan all.
     if FOCUS_MODE:
         ordered = [FOCUS_REPO]

mimo-swarm/scripts/worker-runner.py

@@ -24,6 +24,23 @@ def log(msg):
         f.write(f"[{ts}] {msg}\n")
 
 
+def write_result(worker_id, status, repo=None, issue=None, branch=None, pr=None, error=None):
+    """Write a result file — always, even on failure."""
+    result_file = os.path.join(STATE_DIR, f"result-{worker_id}.json")
+    data = {
+        "status": status,
+        "worker": worker_id,
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+    }
+    if repo: data["repo"] = repo
+    if issue: data["issue"] = int(issue) if str(issue).isdigit() else issue
+    if branch: data["branch"] = branch
+    if pr: data["pr"] = pr
+    if error: data["error"] = error
+    with open(result_file, "w") as f:
+        json.dump(data, f)
+
+
 def get_oldest_prompt():
     """Get the oldest prompt file with file locking (atomic rename)."""
     prompts = sorted(glob.glob(os.path.join(STATE_DIR, "prompt-*.txt")))
@@ -63,6 +80,7 @@ def run_worker(prompt_file):
 
     if not repo or not issue:
         log(f"  SKIPPING: couldn't parse repo/issue from prompt")
+        write_result(worker_id, "parse_error", error="could not parse repo/issue from prompt")
         os.remove(prompt_file)
         return False
 
@@ -79,6 +97,7 @@ def run_worker(prompt_file):
     )
     if result.returncode != 0:
         log(f"  CLONE FAILED: {result.stderr[:200]}")
+        write_result(worker_id, "clone_failed", repo=repo, issue=issue, error=result.stderr[:200])
         os.remove(prompt_file)
         return False
 
@@ -126,6 +145,7 @@ def run_worker(prompt_file):
                 urllib.request.urlopen(req, timeout=10)
             except:
                 pass
+            write_result(worker_id, "abandoned", repo=repo, issue=issue, error="no changes produced")
             if os.path.exists(prompt_file):
                 os.remove(prompt_file)
             return False
@@ -193,17 +213,7 @@ def run_worker(prompt_file):
         pr_num = "?"
 
     # Write result
-    result_file = os.path.join(STATE_DIR, f"result-{worker_id}.json")
-    with open(result_file, "w") as f:
-        json.dump({
-            "status": "completed",
-            "worker": worker_id,
-            "repo": repo,
-            "issue": int(issue) if issue.isdigit() else issue,
-            "branch": branch,
-            "pr": pr_num,
-            "timestamp": datetime.now(timezone.utc).isoformat()
-        }, f)
+    write_result(worker_id, "completed", repo=repo, issue=issue, branch=branch, pr=pr_num)
 
     # Remove prompt
     # Remove prompt file (handles .processing extension)